Broadcom Jericho4 is a 3nm Ethernet router that links over 1M GPUs across data centers with 3.2Tbps ports, lossless RoCE, and MACsec security for AI-scale networking.
Broadcom’s newest Ethernet fabric router, Jericho4, is made to handle the next generation of spread AI infrastructure. As AI models need more computing power than a single data center can offer, Jericho4 offers the high-bandwidth, low-loss infrastructure to link more than a million GPUs in different regions. See how it changes the way AI networks will work in the future.
What Is Broadcom Jericho4
The Broadcom Jericho4 is a cutting-edge Ethernet fabric router that works best in settings with distributed AI computing. It solves a problem that AI faces today: the limited space and power of single data centers. With the rise of models with trillions of parameters and training with billions of tokens, the industry needs compute fabrics that connect computers in different towns, countries, and even continents.
That’s what Broadcom Jericho4 does. Over a million XPUs (accelerators like GPUs or TPUs) can be connected to it, which makes it a key building block for future AI networks.
Featured
Key Features of Broadcom Jericho4
What makes Broadcom Jericho4 different from other networking chips is this:
1. HyperPort Aggregation
Each Jericho4 machine has 36,000 HyperPorts, and each port has 3.2 Tbps of bandwidth because it combines 4 800GE Ethernet links logically. This increases use by 70%, makes planning easier, and makes data transfer very fast.
2. AI-Ready Performance
It was made for lossless RoCE (RDMA over Converged Ethernet), which lets GPUs talk to each other over 100 km or more without any packet loss. This makes it great for AI workloads that need to be sent between data centers.
3. Security at Speed
At full line rate, MACsec encryption can be used on any port. This protects the safety and integrity of data without slowing it down, which is very important for regulated AI and environments with multiple tenants.
4. Built on 3nm
It is made using a 3nm process and has 200G PAM4 SerDes (serializer/deserializer) built in, so it doesn’t need any extra retimers. That means less heat, less power, and a lower cost for the whole system.
5. Interoperability via UEC
It fully complies with the Ultra Ethernet Consortium (UEC), which means it works with a lot of different NICs, switches, and Ethernet-based AI systems.
Why Broadcom Jericho4 Matters for AI Networking
Models like GPT-5 and OpenELM are always pushing the limits of compute, memory, and delay. This means that the infrastructure that supports them needs to change too. Three important problems for large-scale AI activities are solved by Jericho4:
- Scalability Beyond a Single Rack
AI is now too big to fit in one place. Jericho4 allows you to scale compute across multiple physical locations — not just servers. - Bandwidth Without Loss
AI workloads are highly sensitive to packet loss. Deep buffering and intelligent congestion management ensure lossless RDMA even across cities. - Standardized Yet Specialized
By using Ethernet not InfiniBand Jericho4 lowers cost, avoids vendor lock-in, and ensures openness via UEC.
Broadcom Jericho4 vs Nvidia Spectrum-X vs Cisco Silicon One
| Feature | Jericho4 | Spectrum-X | Silicon One |
|---|---|---|---|
| Fabric Type | Proprietary Ethernet | Ethernet | Dual-purpose |
| Port Speed | 3.2 Tbps (HyperPort) | 800G | ~1.6 Tbps |
| Buffering | Deep-buffered | Standard | Mixed |
| Encryption | MACsec at full rate | Optional | Varies |
| Special Cabling | No | No | No |
| Scheduled Fabric | Yes (with Ramon ASIC) | No (Bluefield DPU) | Yes |
| Software | Unified SDK | DPU-powered | IOS-XR or disaggregated |
Who Should Use Jericho4
Jericho4 is the ideal choice for:
- AI labs training multi-trillion parameter models
- Hyperscalers building city-scale AI clusters
- Cloud providers expanding AI-as-a-Service
- Telcos upgrading long-haul networks to AI-grade Ethernet
- Regulated industries needing secure, encrypted data flows
If you’re building for AI at planetary scale, Jericho4 is the core.